Re: A test for replay of regression tests

On Thu, Mar 24, 2022 at 8:56 PM Thomas Munro <thomas(dot)munro(at)gmail(dot)com> wrote:
> Interesting. IIUC your chaos gizmo shows that particular vacuum test
> still failing on master, but that wouldn't happen in real life because
> since 383f2221 it's a temp table. Your gizmo should probably detect
> temp rels, as your comment says. I was sort of thinking that perhaps
> if DISABLE_PAGE_SKIPPING is eventually made to do what its name sounds
> like it does, we could remove TEMP from that test and it'd still pass
> with the gizmo...

Why not just use VACUUM FREEZE? That should work, because it won't
settle for a cleanup lock on any page with an XID < OldestXmin. And
even if there were only LP_DEAD items on a page, that wouldn't matter
either, because we don't need a cleanup lock to get rid of those
anymore. And we consistently do all the same steps for rel truncation
in the no-cleanup-lock path (lazy_scan_noprune) now.

I think that DISABLE_PAGE_SKIPPING isn't appropriate for this kind of
thing. It mostly just makes VACUUM not trust the visibility map, which
isn't going to help. While DISABLE_PAGE_SKIPPING also forces
aggressive mode, that isn't going to help either, unless you somehow
also make sure that FreezeLimit is OldestXmin (e.g. by setting
vacuum_freeze_min_age to 0).

VACUUM FREEZE (without DISABLE_PAGE_SKIPPING) seems like it would do
everything you want, without using a temp table. At least on the
master branch.

--
Peter Geoghegan